Stochastic Nucleation of Polymorphs: Experimental ... - ACS Publications

Jump to Modeling Stochastic Nucleation of Polymorphs - Let us now extend the stochastic model of nucleation to a system with an arbitrary number of ...
0 downloads 0 Views 1MB Size
Article pubs.acs.org/crystal

Stochastic Nucleation of Polymorphs: Experimental Evidence and Mathematical Modeling Giovanni Maria Maggioni, Leonard Bezinge, and Marco Mazzotti* Separation Processes Laboratory, ETH Zurich, Zurich, Switzerland

Crystal Growth & Design 2017.17:6703-6711. Downloaded from pubs.acs.org by IOWA STATE UNIV on 01/18/19. For personal use only.

S Supporting Information *

ABSTRACT: In this work, we investigate the primary nucleation of isonicotinamide, i.e., a compound exhibiting several different polymorphs, from ethanol at several supersaturations and at isothermal conditions. Experiments have been performed over a broad range of supersaturation and in two different volumes, collecting hundreds of data points. These data show that not only the detection times, but also the polymorphs observed in the system in different repetitions of the same experiments are statistically distributed. By using the wealth of experimental data produced and building upon previous theoretical work, we develop a simple stochastic model of primary nucleation from solution for a compound forming an arbitrary number of polymorphs. The model shows that, if the polymorphism of the compound cannot be monitored, the primary nucleation rate estimated from the experiments represents only an apparent nucleation rate. Furthermore, the first polymorph nucleating in the system is not expected to be a function of the system size, but only of its concentration and temperature. The model also allows one to perform a semiquantitative analysis of the system, identifying the regions where each polymorph preferentially forms.

1. INTRODUCTION The formation of a new crystal from a clear solution is an activated, rare event; hence primary nucleation is a stochastic process,1,2 and nucleation times measured in repeated experiments at the same conditions are distributed, e.g., according to a Poisson statistics.3−5 Utilizing experimental protocols that have been discussed in the literature,3,5−7 empirical cumulative distributions of nucleation (or detection) times can be determined and used to estimate the associated nucleation rate. Note that new crystals are detected only after nuclei have evolved into fully grown crystals.8−11 For substances that exhibit different polymorphs, there is the need for establishing which polymorph has been formed and for attributing the estimated nucleation rate to the appropriate polymorph, which is challenging. The Ostwald’s rule of stages, stating that crystals of the least stable polymorph nucleate first and then possibly transform via a solution-mediated mechanism toward more stable forms, was believed to apply,12,13 while it was understood that in a solution at a specific solute concentration the polymorph that forms first is that with the largest nucleation rate among those for which the solution is supersaturated, i.e., often, but not always, the least stable polymorph.14−16 In a deterministic framework, both interpretations take for granted that at a given supersaturation and temperature there is one polymorph that, in repeated nucleation experiments, forms always first. Considering nucleation in a stochastic framework, as we should do, leads to questioning the conclusion above. If © 2017 American Chemical Society

nucleation times are distributed, even if one polymorph has a shorter average nucleation time than another, it might well be that in a specific experiment the nucleation time of the former is longer than that of the latter; hence in that specific experiment, the generally slow-nucleating polymorph nucleates before the generally fast-nucleating polymorph does. Considering that the average nucleation time of a polymorph and the variance of the distribution of nucleation times scale with the reciprocal of the nucleation rate and of the system size, such effects are supposed to be more or less observable depending on the experimental conditions (supersaturation and temperature) and on the size of the crystallizer. Furthermore, it could also happen that the polymorph that forms first transforms into a more stable one before being detected (or before being collected at the end of the experiment for off-line analysis). Experimental evidence concerning the stochastic nature of nucleation of polymorphs has been reported in recent years.17−19 Among others, Kulkarni et al.20,21 and Caridi et al.22 studied the nucleation of isonicotinamide (INA), of which at least three polymorphs are known, in various solvents and solvent mixtures. They reported that when INA is in ethanol (EtOH), methanol (MeOH), and 2-propanol (IPA), only the most stable form (Form II, in their work) nucleated, whereas in nitrobenzene and nitromethane the Form IV and the Form I Received: September 15, 2017 Revised: November 6, 2017 Published: November 8, 2017 6703

DOI: 10.1021/acs.cgd.7b01313 Cryst. Growth Des. 2017, 17, 6703−6711

Crystal Growth & Design

Article

nucleated, respectively. Kulkarni and co-workers8 also studied the crystallization of INA in two solute mixtures: in the case of ethanol/nitrobenzene, they reported the formation of both Form II and Form I, while in the case of ethanol/nitromethane of both Form II and Form V. These authors observed that only one polymorph formed when crystallization was performed in small volumes (3 mL). The polymorph forming in each individual experiment appeared to be random, while the occurrence frequency of the polymorphs depended on the initial concentration. On the contrary, in large volume experiments they observed solid mixtures of multiple polymorphs. In the same work,8 Kulkarni et al. reported that also 4-hydroxyacetophenone in ethyl acetate yielded different polymorphs in different experiments. In this system, though, liquid−liquid phase separation was reported, which may admittedly have affected the nucleation process. Hansen et al.17 investigated the crystallization of a nonsteroidal drug, piroxicam, in acetone and acetone/water mixtures with several additives. They reported that piroxicam crystallized either Form I, or Form II, or a monohydrate crystal, in a random way and with an occurrence frequency depending on the initial solute concentration. To interpret their data, Kulkarni et al.,8 as well as Hansen et al.,17 postulated the stochastic formation from the clear solution of a first single nucleus, dictating the evolution of the system in small volumes. In larger systems, a more conventional polynuclear mechanism would act, resulting in mixtures of multiple solid forms. The formation of different polymorphs at the same conditions randomly, rather than in the deterministic way predicted by the classical theory, has also been recently reported for other systems,5,23,24 such as p-ABA in water/EtOH mixtures,18,19 L-glutamic acid in water,24 and mhydroxibenzoic acid in 1-propanol.25 Furthermore, Cui et al.26 have observed that either form α, or form β, or a mixture of the two crystallized on a functionalized substrate via contact secondary nucleation, which is an activated mechanism27 similar to primary nucleation. It is also worth noting that Hansen et al.28 in a recent study have further investigated the crystallization of INA in different solvents (MeOH, acetonitrile, acetone, ethyl acetate, and dichloromethane) and have concluded that the polymorphism of INA depends not only on the solvent chosen, but also on the concentration and on the temperature at which crystallization is carried out. Recently, the possibility of modeling these phenomena has been demonstrated using an approach based on the description of nucleation as a Poisson process.29,30 By expanding on results that we have presented recently,29 in this work the stochastic nucleation of polymorphs is studied both experimentally (crystallizing INA in ethanol) and theoretically. The paper is structured as follows: experimental methods and experimental results are reported in Sections 2 and 3; the stochastic model is developed, discussed, and utilized in Section 4; conclusions are drawn in Section 5.

correspond to the forms labeled as EHOWIH01, EHOWIH02, and EHOWIH04 in the CSD,20,28 respectively. The solubilities of the polymorphs have been measured gravimetrically, and their values are reported in Table 1. Note that, despite the overlap of the confidence

Table 1. Values of the Solubilities c*j , of the Three Polymorphs of INA in EtOH, Measured Gravimetrically at Different Temperatures [g/kgs] T [K]

cα*

293.15 298.15 303.15 308.15 310.15 316.65

79.02 ± 0.60 93.41 ± 1.00

cβ* 80.69 95.75 115.1 134.4

± ± ± ±

cγ* 3.00 3.20 1.90 3.00

85.22 ± 2.20 100.6 ± 2.20 117.1 ± 4.05

138.4 ± 0.70 170.1 ± 0.30

intervals at 298.15 K, the form α is indeed the most stable, as we could independently verify in ad hoc experiments (see also Section 3.3). We have assumed the ratio of activity coefficients γ(T,c)/γ(T,c*) ≈ 1, and hence the supersaturation Sj for each polymorph has been expressed as c Sj(T , c) = j = α, β, γ c*j (T ) (1) where c is the concentration of INA in EtOH and cj* the solubility of the polymorph j, expressed in grams of solute per kilogram of solvent [g/kgs]. Unless explicitly indicated otherwise, in the following, with the generic term “solubility” and “supersaturation”, we refer to those of the stable polymorph, i.e., cα* and Sα. The crystallization experiments have been performed in a range of supersaturations between Sα = 1.10 and Sα = 2.10. 2.2. Methods. Cooling crystallization experiments of INA in EtOH were performed at a constant temperature of 298.15 K; due to the accuracy of the thermocouples, this temperature as well as the other temperatures reported in the following, are accurate within 0.5 K. The crystallization experiments were performed with an amount of solvent equal either to 1.40 g (small size experiments) or to 50 g (large size experiments), the former in a Crystal16 (Technobis Crystallization Systems, multiple reactor setup with 16 wells, each equipped with an independent thermocouple7), and the latter in an EasyMax 400 (Mettler Toledo). Both devices were equipped with glass reactors; for the Crystal16, standard HPLC glass screw topped vials were used as vessels. As discussed elsewhere, the reactors were properly sealed during the experiments to minimize solvent evaporation, since this would change the solute concentration and hence the supersaturation.7 We have also monitored the evaporation by weighing the reactors (individual vials in Crystal16 and single glass reactor in EasyMax) before and after each experiment: if evaporation resulted in a change of supersaturation larger than 2%, the experiment was discarded and repeated. The vials in the Crystal16 were stirred with magnetic bars at 700 rpm, while the reactor in the EasyMax with a 4-blade impeller at 400 rpm, to guarantee that the system was well-mixed. The stock solution was kept for 60 min at 333 K to ensure the complete dissolution of INA and immediately thereafter used to fill the vials. For the large scale experiments, the solution was prepared directly inside the glass reactor. Supersaturation was induced by cooling from different saturation points, and multiple cycles were performed for both small and large size experiments. A crystallization cycle consisted of four stages: a heating ramp from 298.15 to 333.15 K (2.0 and 2.5 min, for Crystal16 and EasyMax, respectively), a dissolution step at 333.15 K (60 min, both devices), a cooling ramp from 333.15 to 298.15 K (2.0 and 2.5 min, for Crystal16 and EasyMax, respectively), and finally an isothermal step at 298.15 K (its duration varied from 10 to 180 min, depending on the supersaturation and the system size). The differences in the heating and cooling stage depended on the different configuration and controller of the devices used.

2. EXPERIMENTAL SECTION 2.1. Materials. Isonicotinamide (INA), purity ≥99%, was purchased from Sigma-Aldrich; ethanol (EtOH) from Fluka (Ethanol Laboratory Reagent, absolute, ≤99.5%) and EMD Millipore (Ethanol Absolute For Analysis Emsure Acs, Iso, Reag. Ph. Eur.). Prior to crystallization experiments, the solution has been filtered through 13 mm syringe filters (PTFE hydrophobic, nonsterile, pore size 0.22 μm); INA was used as received. Different polymorphs of INA are known, and their nomenclature in the literature20,28 appears somewhat confusing. In this work, we have labeled the three polymorphs in order of decreasing stability at 298 K in EtOH as α, β, and γ. They 6704

DOI: 10.1021/acs.cgd.7b01313 Cryst. Growth Des. 2017, 17, 6703−6711

Crystal Growth & Design

Article

Figure 1. Characterization of the polymorphs. (a) The XRD spectra of the powder of α, β, and γ; the bands in blue, magenta, and yellow, highlight the characteristic peaks of the three polymorphs, respectively. (b) The Raman spectra of pure α, β, and γ in ethanol, in blue, red, and yellow, respectively. 2.3. Measurement Protocols. The polymorphism of the crystals produced in the experiments was assessed off-line, ex situ through Xray diffraction (XRD) measurements (experiments in Crystal16 and EasyMax) and online, in situ through Raman spectroscopy (RA 400 Raman spectrometer from Mettler Toledo, for the experiments in the EasyMax). XRD and Raman spectra of the pure polymorphs are shown in Figure 1, panels a and b, respectively. Three characteristic peaks for each polymorph, i.e., peaks which are exhibited by that polymorph but not by the others, have been highlighted in Figure 1a as vertical bands in blue (2θ = 17.8, 26.0, 31.0), red (2θ = 14.5, 23.1, 32.2), and yellow (2θ = 23.9, 26.9, 37.0), for α, β, and γ, respectively. Note that the peak at 2θ = 23.4, though very prominent for polymorphs α and γ, is not selected because of its ambiguous attribution. For the experiments in the Crystal16, we measured the XRD spectra only of the crystals produced during the very last cycle of a series of experiments using the same solution. The Raman spectra of the pure forms in EtOH are reported in Figure 1b and are consistent with those given by Kulkarni et al.20 The three polymorphs can be clearly distinguished in two intervals of the Raman spectrum, namely, 980−1010 cm−1 and 1200−1230 cm−1, as observed in Figure 1b, where the characteristic peaks of α, β, and γ are highlighted by the blue, red, and yellow bars. Raman spectra were collected in the range of Raman shift from 100 to 3600 cm−1, with a resolution of 1 cm−1 and averaged over 10 scans using an exposure time of 6 s, corresponding to one measurement per minute. The raw data from Raman measurements have been treated according to standard procedures for spectroscopic data pre- and postprocessing: the signal has been smoothed with a Savitzky-Golay filter,31 and a linear baseline correction32 has been applied to the spectra. In the EasyMax, at least nine cycles for each supersaturation Sα have been performed, thus obtaining a number of repetitions comparable with those (16) obtained from the last cycle of the experiments in the Crystal16. We used Raman spectroscopy also to verify that the system attained unbiased initial conditions during the dissolution step after each crystallization step. The complete disappearance of the peaks associated with the solid forms of INA within 5 min from the beginning of the dissolution stage and the reproducibility of the Raman spectra of the clear solution at high temperature in different cycles indicated that the initial conditions of the system were the same at the beginning of each crystallization experiment. Several XRD spectra measured from samples of small scale experiments were markedly different from those of pure polymorphs. Since solvates of INA have been reported for crystallization from water,33 we have taken care of excluding the presence of solvates or solvent inclusions by means of two different tests. The first test consisted of repeating the XRD measurements on the same samples

after 5 days, during which the crystals were kept in a dry environment: all spectra were unchanged. The second test was a thermal gravimetric measurement, with temperatures up to 500 K, i.e., well beyond the melting temperature of any possible solvent, of all known polymorphs and of any solvates. The thermal gravimetric analysis showed no loss of material, hence indicating that no ethanol was trapped in the solids and also pointing out that the spectra exhibiting characteristic peaks of two or three polymorphs were originated by a solid mixture of pure crystals (i.e., a mixture of polymorphs). Detection times, tD, were also recorded during all experiments (see Figure 2). For the Crystal16 experiments, the detection time was

Figure 2. An illustration of detection times based on the temperature profile (top) and on the Raman intensity I (bottom), measured for a peak typical of ethanol, for the case with S = 1.30. T* indicates the saturation temperature of the most stable form. defined as the difference in time between the attainment of the isothermal temperature and the decrease of light transmissivity below a given threshold. Following the same protocol discussed in detail elsewhere,7 we retained only the detection times in the small scale experiments whose temperature was within a band of ±0.5 K from the set value of 298.15 K. For the EasyMax experiments, the detection time was defined as the difference in time between the attainment of the saturation point and the time when a characteristic peak of ethanol (at wavelength 883 6705

DOI: 10.1021/acs.cgd.7b01313 Cryst. Growth Des. 2017, 17, 6703−6711

Crystal Growth & Design

Article

cm−1) decreased below a threshold value with respect to the peak maximum in the dissolved state before cooling (see Figure 2). This unconventional choice of defining the detection condition with respect to a property of the liquid phase, rather than to one of the solid, has been based on the fact that the intensity of this ethanol peak was always quite strong and clearly weaker upon the formation of crystals. On the contrary, the peaks associated with the solid forms of INA were initially quite weak at low supersaturations, becoming more visible only upon significant crystal growth. For this reason, we opted to define detection with respect to the ethanol peak at 883 cm−1. We note that nucleation (and detection) can occur also during the cooling ramp, as soon as Sj > 1: this can be more likely observed in systems at high supersaturations and/or of large size, since the average nucleation time is inversely proportional to the system size and the nucleation rate, the latter increasing with increasing Sα.4,10 Indeed, in the large size experiments detection occurred always during the cooling ramp for all experiments with Sα > 1.40. A direct quantitative comparison between the values of the detection times measured in the Crystal16 and those measured in the EasyMax is not possible, since different detection techniques and detection conditions were applied in the two devices.

detection times (n ≤ N, since not all runs produce crystals during the duration of the experimental observation), with which we form the ordered vector tD: tD = [tD ,1, tD ,2 , ..., tD , n]T : 0 ≤ tD ,1 ≤ tD ,2 ≤ ... ≤ tD , n < +∞ (2) through which the following empirical cumulative distribution function (eCDF) F*N (tD) is defined as FN*(t D) =

1 N

n

∑ [tD,j , +∞)(tD) j=1

(3)

where [a , +∞) is the indicator function on the interval [a, +∞), i.e., equal to 1 if a ≤ tD < ∞, and 0 otherwise. The vector tD constitutes a discrete, finite sample of the stochastic variable tD, whereas FN*(tD) is an estimate of F(tD). 3.2. Polymorphic form. Figure 4a (Crystal16) and Figure 4b (EasyMax) demonstrate that in repeated experiments carried out at

3. EXPERIMENTAL RESULTS We have performed two sets of nucleation experiments, which provide different, complementary pieces of information. The Crystal16 allows one, on the one hand, to gather a large number of detection time data, but cannot monitor which polymorph is associated with such events, and the XRD measurements can be obtained only from the samples collected after the final cycle of a series. On the other hand, the EasyMax has a much lower productivity in terms of number of experiments, but it allows one to measure online the time-resolved Raman spectra. 3.1. Detection Times. Figure 3a (Crystal16) and 3b (EasyMax) illustrate for each experimental supersaturation (horizontal axis) the

Figure 4. Evidence of polymorph formation in Crystal16 (a) and in EasyMax (b). (a) XRD spectra at Sα = 2.00 obtained from four different repetitions (labeled as E1, E2, E3, E4) of the same typical nucleation experiment. Depending on the vial analyzed, either one of the three forms (or possibly a mixture) is observed. The colors of the vertical bands associated with the peaks typical of each polymorph, as in Figure 1: blue, red, and yellow refer to α, β, and γ, respectively. (b) Relative occurrence of the three polymorphs in the experiments carried out at different supersaturations. Blue, red, and yellow bars refer to α, β, and γ polymorphs. Mixed-color striped bars in panel b indicate the experiments in which two different polymorphs were observed together. The thicker bars indicate the relative occurrence at detection (see the detection times in Supporting Information), while the thinner bars with lighter colors indicate the relative occurrence at the end of the experiment.

Figure 3. Detection times measured from the experiments in the Crystal16 (a) and in the EasyMax (b) at different supersaturation levels. M is the mass of ethanol used in the experiments.

exactly the same conditions from identical initial solutions, different polymorphs are obtained first. This is illustrated in Figure 4a through XRD spectra of the final crystals obtained in four different experiments at the same supersaturation: for each experiment, we have indicated the polymorph to which the crystal powder has been attributed. The XRD spectrum of experiment E4 clearly belongs to the α polymorph. The characteristic peaks of the β polymorph are present in the XRD spectra of experiments E1, E2, and E3, although with different intensities. Experiments E1 and E3 also present two peaks characteristic of γ, while E2 two characteristic peaks of α, i.e., experiments E1, E2, and E3, seem to be mixtures of different polymorphs. However, it is evident that the solid products obtained in different experiments are different, even though they have been produced at the same conditions and from the same stock solution. The samples exhibiting mixed XRD spectra, such as those of E1, E2, and E3 in Figure 4a, were analyzed again after 3 and 6 days; the spectra did not change, further confirming that the product crystals did not contain any trace of solvent. As mentioned in Section 2.3, we have also verified by thermal gravimetric analysis that the extra peaks could not be related to solvates. This analysis is also confirmed by observing the

detection times measured in tens or hundreds of repeated experiments (vertical axis). The detection times are clearly distributed, and their spread reflects the stochastic nature of primary nucleation, which has already been reported for several systems.3,4,9,34−36 In Figure 3a, tD varies from a minimum of about 1 min to a maximum of about 5 h at Sα = 1.20, and from a minimum of about 1 min to a maximum of about 30 min at Sα = 2.10. In Figure 3b, for the same supersaturation values, the maximum tD decreases to about 10 and 4 min, at Sα = 1.20 and Sα = 2.10, respectively. This sharp decrease of the broadness of the distribution observed in the data when increasing the supersaturation, from Sα = 1.20 to Sα = 2.08, or the system size, from 1.40 to 50 g of solvent, is also consistent with the statistics associated with the measurement of stochastic processes.3,4 Since nucleation is stochastic, the detection time tD is a stochastic variable (as illustrated in Figure 3), distributed according to an intrinsic underlying cumulative probability function, F(tD). For each supersaturation, we have performed N experiments and measured n 6706

DOI: 10.1021/acs.cgd.7b01313 Cryst. Growth Des. 2017, 17, 6703−6711

Crystal Growth & Design

Article

Figure 5. (a) The empirical cumulative distribution functions for experiments at S = 1.30, 1.58, and 1.91 in the Crystal16; the symbols represent the samples for which the XRD spectra have been measured. The blue, red, and yellow symbols refer to α, β, and γ, respectively. (b) The empirical cumulative distribution functions for S = 1.30 and 2.10 obtained from identifying the polymorph nucleating with time-resolved Raman spectroscopy in the EasyMax, where a much smaller number of points was available. The colored empirical cumulative distributions represent the frequency of the individual polymorph (see also Section 4.1). XRD spectra of the powder obtained from experiments at other supersaturations: examples of such XRD spectra at four different supersaturations are reported in Supporting Information. Figure 4b shows at each supersaturation the relative occurrence Q of the three polymorphs (i.e., the number of experiments in which a specific polymorph was observed, as a percentage of all experiments performed at that specific supersaturation level), as obtained from the Raman spectra both at detection (large vertical bars) and at the end of the experiment (thin vertical bars). The Raman spectra of the experiments performed at four supersaturations, used to construct Figure 4b, are reported in Supporting Information. The mixed-color striped bars in Figure 4b indicate the experiments in which two different polymorphs were observed together. Depending on the supersaturation, between 10% (e.g., Sα = 1.10 and 1.20) and 50% (e.g., Sα = 1.50) of the experiments exhibit mixed Raman spectra, which are associated with the presence of multiple polymorphs. The β and γ forms appear typically together for Sα > 1.50, whereas the α form appears with β for the lowest supersaturation, Sα = 1.10. Two remarks are worth making. First, the distribution of relative occurrences changes markedly from one supersaturation to the next; we will provide a justification thereof in Section 4, based on the different supersaturation dependence of the nucleation rate of the different polymorphs. Second, the distribution of relative occurrences changes from the time of detection to the end of the experiment, because of polymorph transformation, as discussed in Section 4. Figure 5 illustrates the distribution of detection times as observed in the same experiments, by plotting the corresponding empirical cumulative distribution functions (eCDFs). Figure 5a shows the eCDFs (as black lines) of the data obtained in the Crystal16 at Sα = 1.30, 1.58, and 1.91, as computed using eq 3; for each supersaturation level, the experiments for which the XRD spectra were also measured are represented with a colored symbol, in blue, red, or yellow to indicate that polymorphs α, β, or γ, respectively, has nucleated first. It is apparent that all polymorphs have distributed detection times, and any among the three can form first. Figure 5b shows the eCDFs of the data measured in the EasyMax experiments at Sα = 1.30 and 2.10, after analyzing the Raman spectra at detection. The black eCDF is that of the detection times, when ignoring which specific polymorph was detected, i.e., the same type of curve plotted in Figure 5a (but with a much smaller number of data points). The colored eCDFs are those of the individual polymorphs, where the horizontal coordinate, i.e., the detection time, is the same as that of the corresponding point on the black eCDF, whereas the vertical coordinate, corresponding to the cumulative probability, attains a

smaller value because it is calculated on the data points of that polymorph only. 3.3. Solid Phase Transformation. Figure 4b provides evidence that new crystals, after forming in a certain polymorphic form, whose relative occurrence depends on supersaturation, may transform into a more stable form, possibly through a solvent mediated transformation that has been well documented by several other studies,18,20 including a few from our group.37,38 While Figure 4b provides information about the polymorphic form at two points in time, namely, detection and experiment end, and proves that less stable forms may evolve into more stable ones, time-resolved Raman spectra, as those shown in Figure 6, allow monitor of the dynamics of the whole transformation with great clarity.

Figure 6. Transformation of γ polymorph, monitored over 30 min. (a) A direct transformation of γ into α at low supersaturations. (b) A transformation from γ into β, at high supersaturations. Typically, the β form is then stable for several hours, but transforms ultimately into the α form. More specifically, Figure 6a shows the transformation from polymorph γ to α occurring at Sα = 1.30, and Figure 6b shows that from polymorph γ to β occurring at Sα = 1.90; in both cases seven spectra are shown, namely, from detection to 30 min after that, with each spectra taken 5 min after the previous. A few remarks are worth making. First, the transformation in Figure 6a is possibly not yet completed, as the characteristic peak of polymorph γ is still visible in the last spectra recorded. This indicates that at these conditions the γ to α transformation takes more than 30 min. Second, in the case of Figure 6b the transformation of form β to form α is not visible, but it indeed occurred as we could observe after 6 h during which the suspension of β crystals had been left undisturbed. 6707

DOI: 10.1021/acs.cgd.7b01313 Cryst. Growth Des. 2017, 17, 6703−6711

Crystal Growth & Design

Article

Incidentally, we note that the transformation from the β to the α form also proves that α is the most stable form observed in our experiments. Finally, as we observe with reference to the experiment in Figure 6b that the transformation from the β to the α form may take very long, we also conclude that when the most stable polymorph α is present at detection (see measurements at low supersaturation levels in Figure 4b), it has most likely formed via primary nucleation and not through solid phase transformation from another metastable polymorph (which would have taken too long to occur in the short time before detection). 3.4. Discussion of Experimental Results. The experiments in the Crystal16 and in the EasyMax provide complementary information and allow one o identify the main features of the nucleation process, particularly the effect of the system volume and of the solute concentration. The experimental data clearly show not only that detection times are statistically distributed (see Section 3.1, Figure 3), but also that the type of polymorph nucleating in each experiment is itself a stochastic property (Section 3.2, Figure 4). Moreover, the data indicate that the relative occurrence of the different polymorphs depends strongly on the solute concentration (see Figure 5b). Therefore, a proper understanding of this nondeterministic behavior requires analyzing the experiments in the context of statistical models of nucleation.3,34,35,39 Stochastic models of nucleation for nonpolymorphic systems describe the nucleation time, tN, i.e., the time when the first primary nucleus forms, with S > 1, as a stochastic quantity. Even though in some cases crystal growth has been found to be itself stochastic,40−42 primary nucleation is considered to be the major source of stochasticity during crystallization from a clear solution, whereas growth can be considered to be deterministic and can be modeled accordingly. Unfortunately, nucleation times tN cannot be measured directly, because the nuclei are too small for being observed.3,4,9−11,35 Consequently, one measures tD and, using a model, retrieves tN, from which then the nucleation kinetics can be estimated. Since nucleation times are distributed,3,9,34 also detection times are distributed. A thorough and comprehensive discussion about the estimation of tN from tD and the necessary assumptions behind it can be found elsewhere.3,6,9−11

represents a possible state different from the empty state. The (random) appearance of the first nucleus causes the irreversible transition of the system from the original empty state to the new state. The functions P0(t) and Pj(t) represent the probabilities that, at time t, the system is still in the initial state and that it is in the state j, respectively. Note that the system must be in one of the possible states, therefore π

P0(t ) +

∑ Pj(t ) = 1 (4)

j=α

and that the probability P(t) that a nucleus of any polymorph has formed, irrespective of which one, is by definition of P0(t) simply: π

P(t ) = 1 − P0(t ) =

∑ Pj(t ) (5)

j=α

The evolution of P0(t) and Pj(t) can be described in terms of the rates of transition between the possible states of the system. The rate of change of the probability Pj(t) (state with one nucleus of polymorph j) is given by the probability P0(t) (state without nuclei) times the transition rate λj, while the rate of change of P0 is minus the product of P0(t) times an overall transition rate λ̂: ⎧ dP0 = −λ P̂ 0 ⎪ ⎪ dt ⎨ ⎪ dPj = λjP0 ∀ j = α , ..., π ⎪ ⎩ dt

(6) (7)

The corresponding initial conditions are ⎧ ⎪ P0(0) = 1 ⎨ ⎪ ⎩ Pj(0) = 0

4. MODELING STOCHASTIC NUCLEATION OF POLYMORPHS 4.1. Mathematical model. Primary nucleation has successfully been modeled as a Poisson process,3,4,9−11,21,35,36,43 ,43 which is the mathematical formalism describing the occurrence of rare events,44 e.g., activated phenomena such as chemical reactions and nucleation.2 It is often assumed that only the first nucleus forms stochastically at the time tN, whereas the later nucleiprimary and secondary alikeform deterministically.3,4,6,10,11,35 For the sake of simplicity, but without loss of generality, we consider here only unseeded isothermal crystallization.4,10,35 Note that the primary nucleation rate J is constant in an isothermal, closed system with no crystal and no chemical reactions, since the solution composition does not change until the first nucleus forms. We also emphasize that assuming the first nucleation event as the only source of stochasticity is a rather strong hypothesis, whose validity for nonpolymorphic systems we discussed in another work.10 This hypothesis, however, allows one to build a model conceptually simple, computationally affordable, and semiquantitatively accurate. Let us now extend the stochastic model of nucleation to a system with an arbitrary number of polymorphs (α, ..., π). Such a system can be thought of as occupying discrete states: an empty state with no nuclei (which is also the initial state of the system) and a state where a nucleus of polymorph j has formed (j = α, ..., π); hence, the presence of a nucleus of polymorph j

(8) ∀ j = α , ..., π

(9)

By summing term by term all eqs 6 and 7 and employing the conservation of total probability given by eq 4, one obtains: π

λ̂ =

∑ λj (10)

j=α

which is a physically sensible result. On the basis of physical arguments,35 λj is the product of the primary nucleation rate of the polymorph j, Jj, and the system size, V, i.e., λj = JjV (j = α, ..., π); hence λ̂ = V∑j π= αJj = V J,̂ where J ̂ = ∑j π= αJj represents an apparent nucleation rate and corresponds to what an observer incapable of distinguishing the different polymorphs would measure. Solving eqs 6 and 7, using also eq 5, yields P0(t ) = exp( −λ t̂ )

(11)

P(t ) = 1 − exp( −λ t̂ )

(12)

λj P(t ) λ̂

(13)

Pj(t ) =

∀ j = α , ..., π

Note that Pj(t) is the so-called joint probability, i.e., the probability that at time t a nucleus has formed and it is of polymorph j. Dividing Pj by the probability P that a nucleus of any type has formed, yields the occurrence probability ξj: 6708

DOI: 10.1021/acs.cgd.7b01313 Cryst. Growth Des. 2017, 17, 6703−6711

Crystal Growth & Design ξj =

Pj(t ) P(t )

=

Jj λj = Ĵ λ̂

∀ j = α , ..., π

Article

depends on the conditions of the system, but not on the system size. Finally, we can use the model to investigate the influence of the operating parameters, namely, the solute concentration and the system temperature. 4.2. Comparison between the Model Results and the Experiments. In this section, we focus on studying the effect of the solute concentration on the system; since experiments were isothermal, we equivalently look at the influence of the supersaturation, Sα. The primary nucleation rate of the individual polymorphs is assumed to follow the expression of the Classical Nucleation Theory:1,45,46

(14)

which is the conditional probability, namely, the probability that, at any time, the nucleus which has formed belongs to polymorph j, given that a (first) nucleus of any type has formed. Note also that, for a homogeneous system such as the one considered here, ξj is a function only of the nucleation kinetics. Figure 7 illustrates the results obtained above for the case with three polymorphs, namely, j = α, β, γ. The black solid line

⎛ ⎞ Bj ⎟ Jj (c , T ) = Sj(c , T )Aj (T ) exp⎜⎜ − 3 2 ⎟ ⎝ T ln Sj(c , T ) ⎠ j = α, β, γ

(15)

At constant temperature T, the pseudorate J ̂ is then a function of the actual solute concentration c (through the supersaturation, eq 1) and of all individual nucleation rates of α, β, and γ: J (̂ c) = Jα (c) + Jβ (c) + Jγ (c)

(16)

Equations 15 and 16 show that such model needs six parameters to describe the INA/EtOH isothermal system, namely, the pair (Aj, Bj) for each polymorph, with j = α, β, γ. Determining these six parameters is not trivial; two points are worth mentioning. First, when no information about the specific polymorph forming is available, such as during detection time experiments in the Crystal16, the empirical cumulative distribution function obtained from experiments allows estimating only the value of J ̂ at the chosen supersaturation level. Second, a proper estimation of the parameters would require the characterization not only of crystal growth, secondary nucleation, and dissolution, but also of the detection technique itself:11,47 such studies go beyond the scope of this contribution. For these reasons, even though we have selected the parameters (see Table 2) to run our

Figure 7. Cumulative probabilities for a system with three polymorphs in a system with V = 1.8 mL. The black, blue, red, and yellow lines refer to the probability of forming any polymorph, α, β, γ, respectively, and are calculated for the nucleation rates J ̂ = 490, Jα = 32, Jβ = 48, Jγ = 410 [#/s/m3]. The values of the nucleation rates have been computed using eqs 15 and 16 with the parameters reported in Table 2, at T = 298.15 K and Sα = 1.50.

represents the cumulative probability function given by eq 12, while the blue, magenta, and yellow line, for form α, β, and γ, respectively represent the cumulative probability functions given by eq 13. The probability P, associated with J,̂ behaves identically to the cumulative probability of a standard isothermal Poisson process and, as expected, asymptotically approaches 1 for t → ∞. On the contrary, t → ∞ the probabilities Pj(t) do not approach 1, but the asymptotic value ξj, thus reflecting the fact that a polymorph can or cannot form. Figure 7 also shows clearly that, as long as the value of J ̂ is constant (see eq 14), the occurrence probability of the polymorphs does not depend on the nucleation time itself, as indicated by eq 14: to change such probability, one must alter the nucleation kinetics of the polymorphs, for instance, by operating at a different supersaturation, or at a different temperature. It is worth highlighting three points. First, according to eq 12 all polymorphs contribute to the total nucleation probability, P(t), through their individual intensities λj, hence ultimately through Jj. Second, the polymorphs associated with higher nucleation rates are more likely to nucleate, but in general they do not necessarily nucleate first in each individual repetition of an experiment: all polymorphs have nonzero (albeit possibly very small) probability to nucleate, unless their supersaturation is below 1. Third, the system size V influences the nucleation probability P, i.e., the nucleation time, but does not affect the occurrence probability ξj. Thus, which polymorph nucleates

Table 2. Parameters of the Nucleation Rate Jj Selected and Used in the Simulations parameters/form

α

β

γ

Aj [#/s/m3] Bj [K3]

23 475

17715 2.37 × 107

392 8.85 × 105

simulations based on a fitting procedure, we consider such values appropriate only for semiquantitative studies. The selection procedure is detailed in Supporting Information. Let us now compare the simulations results and the experimental data. Equation 15 indicates that one can modify the occurrence probability of the polymorphs (eq 14) by varying the supersaturation, hence the nucleation rate; using the parameters reported in Table 2, one can compute the values of ξα, ξβ, and ξγ as a function of Sα. These values are represented in Figure 8 as the blue, red, and yellow solid line, for ξα, ξβ, and ξγ, respectively. Figure 4b, on the other hand, reports the (percentage) occurrence frequency Q of form α, β, and γ, at different values of Sα: one can interpret the occurrence frequency of each polymorph as an estimate of its occurrence probability. By normalizing Q from Figure 4b, one can then compare the values of Q with those of ξj, at each Sα, as shown in Figure 8. 6709

DOI: 10.1021/acs.cgd.7b01313 Cryst. Growth Des. 2017, 17, 6703−6711

Crystal Growth & Design

Article

Thus, summarizing, in this work we have presented nucleation experiments of isonicotinamide in ethanol where all polymorphs, and not only the stable one (α, form EHOWIH01), have been crystallized from clear solutions. The production of the metastable β and γ forms (form EHOWIH02 and EHOWIH04, respectively) has been consistently obtained in a broad range of initial concentrations, and has been shown to occur also at two different system sizes. The formation of the polymorphs β and γ is in contrast with one hypothesis made in the literature about crystallization of INA in EtOH,20,22 but is consistent with the behavior of INA observed in other solvents.28 Even more interestingly, our experiments exhibited a stochastic nature in terms not only of detection times, but also of the polymorph forming first. On the basis of the theoretical results concerning the stochastic nature of primary nucleation, we have developed (Section 4) a simple model which naturally extends previous work on stochastic primary nucleation4,9−11,35 to a polymorphic system. Such model describes in a coherent manner the stochasticity of nucleation/detection times, as well as that of the type of polymorph nucleating, as observed in the experiments.

Figure 8. Occurrence probabilities ξα, ξβ, and ξγ, in blue, red, and yellow, plotted as a function of Sα, while the stacked bars represent the rescaled relative occurrences Q measured in the large scale experiments and reported also in Figure 4b.



One can readily see that the model results and the experimental measurements exhibit a very similar behavior. Thus, concluding, these results suggest that the model presented in Section 4.1 is not only plausible, but also consistent with the data, further indicating that it indeed describes the main, fundamental physical features of the system.

ASSOCIATED CONTENT

S Supporting Information *

The Supporting Information is available free of charge on the ACS Publications website at DOI: 10.1021/acs.cgd.7b01313. Two sections: in the first one (Appendix A), we have detailed how the parameters for the simulations in Section 4 have been selected; in the second one (Appendix B), we have reported additional experimental data mentioned in Section 2, namely, further XRD spectra and further Raman spectra, and the table reporting the detection times associated with each Raman spectrum measured (PDF)

5. DISCUSSION AND CONCLUSIONS First, we make a few remarks about the deterministic, kinetic interpretation of Ostwald’s rule of stages developed by Davey, Garside, Cardew, and others,12−16,48 mentioned in Section 1. Even though the deterministic theory cannot explain the experiments of Section 3 of INA in EtOH, a deep connection exists between such theory and the stochastic model. Davey, Garside, and others interpreted the primary nucleation rate as a measure of how much a certain polymorph j is favored against the others in a deterministic framework. This “bias” toward a specific polymorph would promote the formation of precritical clusters of a specific configuration and would eventually lead to the formation of the polymorph with the highest nucleation rate at the given conditions. In the stochastic framework, the “bias” of the deterministic theory can be interpreted as the probability of the system to form a certain polymorph (see eq 14). In fact, the intervals of Sα where the stochastic model predicts that a polymorph j is more likely to appear than the others correspond to the intervals where the deterministic theory predicts that the same polymorph j nucleates always first. Thus, the deterministic model can be seen as a special case of the stochastic model, in the sense that its prediction corresponds to the most-likely outcome predicted by the stochastic approach. Second, we note that when Jj at the given conditions is such that ξj ≈ 1, then the system exhibits a pseudodeterministic behavior; i.e., the experiments yield almost always the polymorph j only. Finally, it is worth noting that, through a qualitative analysis not reported here, we can show that the region where the α polymorph nucleates with higher probability should become larger when T decreases; this conclusion further supports the conjecture that the difference between our findings and those of Kulkarni et al.20 can be due to the different temperatures at which these authors have conducted their experiments, namely, 298.15 K in our case and 278.15 K in theirs.



AUTHOR INFORMATION

Corresponding Author

*E-mail: [email protected]. Phone: +41 44 632 24 56. Fax: +41 44 632 11 41. ORCID

Marco Mazzotti: 0000-0002-4948-6705 Notes

The authors declare no competing financial interest.



REFERENCES

(1) Kashchiev, D. Nucleation, Basic Theory with Applications; Butterworth Heinemann, 2000. (2) Davey, R. J.; Schroeder, S. L. M.; ter Horst, J. H. Angew. Chem., Int. Ed. 2013, 52, 2166−2179. (3) Jiang, S.; ter Horst, J. H. Cryst. Growth Des. 2011, 11, 256−261. (4) Kadam, S. S.; Kulkarni, S. A.; Coloma Ribera, R.; Stankiewicz, A. I.; ter Horst, J. H.; Kramer, H. J. Chem. Eng. Sci. 2012, 72, 10−19. (5) Little, L. J.; Sear, R. P.; Keddie, J. L. Cryst. Growth Des. 2015, 15, 5345−5354. (6) Xiao, Y.; Tang, S. K.; Hao, H.; Davey, R. J.; Vetter, T. Cryst. Growth Des. 2017, 17, 2852−2863. (7) Maggioni, G. M.; Bosetti, L.; dos Santos, E.; Mazzotti, M. Crys. Growth Des. 2017, in press. (8) Kulkarni, S. A.; Meekes, H.; ter Horst, J. H. Cryst. Growth Des. 2014, 14, 1493−1499. (9) Sullivan, R. A.; Davey, R. J.; Sadiq, G.; Dent, G.; Back, K. R.; ter Horst, J. H.; Toroz, D.; Hammond, R. B. Cryst. Growth Des. 2014, 14, 2689−2696.

6710

DOI: 10.1021/acs.cgd.7b01313 Cryst. Growth Des. 2017, 17, 6703−6711

Crystal Growth & Design

Article

(47) Nagy, Z. K.; Fujiwara, M.; Woo, X. Y.; Braatz, R. D. Ind. Eng. Chem. Res. 2008, 47, 1245−1252. (48) Bernstein, J.; Davey, R. J.; Henck, J.-o. Angew. Chem., Int. Ed. 1999, 38, 3440−3461.

(10) Maggioni, G. M.; Mazzotti, M. Faraday Discuss. 2015, 179, 359− 382. (11) Maggioni, G. M.; Mazzotti, M. Cryst. Growth Des. 2017, 17, 3625−3635. (12) Ostwald, W. Z. Phys. Chem. 1897, 22, 289−302. (13) Nývlt, J. Cryst. Res. Technol. 1995, 30, 443−449. (14) Cardew, P. T.; Davey, R. J.; Ruddick, A. J. J. Chem. Soc., Faraday Trans. 2 1984, 80, 659−668. (15) Davey, R.; Cardew, P.; McEwan, D.; Sadler, D. J. Cryst. Growth 1986, 79, 648−653. (16) Davey, R.; Garside, J. From Molecules to Crystallizers; 2000; pp 1−52. (17) Hansen, T. B.; Qu, H. Cryst. Growth Des. 2015, 15, 4694−4700. (18) Black, J. F. B.; Davey, R. J.; Gowers, R. J.; Yeoh, A. CrystEngComm 2015, 17, 5139−5142. (19) Garg, R. K.; Sarkar, D. J. Cryst. Growth 2016, 454, 180−185. (20) Kulkarni, S. A.; McGarrity, E. S.; Meekes, H.; ter Horst, J. H. Chem. Commun. 2012, 48, 4983. (21) Kulkarni, S. A.; Kadam, S. S.; Meekes, H.; Stankiewicz, A. I.; ter Horst, J. H. Cryst. Growth Des. 2013, 13, 2435−2440. (22) Caridi, A.; Kulkarni, S. A.; Di Profio, G.; Curcio, E.; Ter Horst, J. H. Cryst. Growth Des. 2014, 14, 1135−1141. (23) Bhamidi, V.; Lee, S. H.; He, G.; Chow, P. S.; Tan, R. B. H.; Zukoski, C. F.; Kenis, P. J. A. Cryst. Growth Des. 2015, 15, 3299−3306. (24) Jiang, N.; Wang, Z.; Dang, L.; Wei, H. J. Cryst. Growth 2016, 446, 68−73. (25) Liu, J.; Svärd, M.; Rasmuson, Å. C. Cryst. Growth Des. 2014, 14, 5521−5531. (26) Cui, Y.; Stojakovic, J.; Kijima, H.; Myerson, A. S. Cryst. Growth Des. 2016, 16, 6131−6138. (27) Agrawal, S. G.; Paterson, A. H. J. Chem. Eng. Commun. 2015, 202, 698−706. (28) Hansen, T. B.; Taris, A.; Rong, B.-G.; Grosso, M.; Qu, H. J. Cryst. Growth 2016, 450, 81−90. (29) Maggioni, G. M.; Mazzotti, M. 48th BACG Annual Conference, Manchester, June 2017. (30) Sun, W.; Ceder, G. CrystEngComm 2017, 19, 4576−4585. (31) Savitzky, A.; Golay, M. J. E. Anal. Chem. 1964, 36, 1627−1639. (32) Mazet, V.; Carteret, C.; Brie, D.; Idier, J.; Humbert, B. Chemom. Intell. Lab. Syst. 2005, 76, 121−133. (33) Báthori, N. B.; Lemmerer, A.; Venter, G. A.; Bourne, S. A.; Caira, M. R. Cryst. Growth Des. 2011, 11, 75−87. (34) Izmailov, A. F.; Myerson, A. S.; Arnold, S. J. Cryst. Growth 1999, 196, 234−242. (35) Goh, L.; Chen, K.; Bhamidi, V.; He, G.; Kee, N. C. S.; Kenis, P. J. A.; Zukoski, C. F.; Braatz, R. Cryst. Growth Des. 2010, 10, 2515− 2521. (36) Brandel, C.; ter Horst, J. H. Faraday Discuss. 2015, 179, 199− 214. (37) Cornel, J.; Lindenberg, C.; Mazzotti, M. Cryst. Growth Des. 2009, 9, 243−252. (38) Cornel, J.; Kidambi, P.; Mazzotti, M. Ind. Eng. Chem. Res. 2010, 49, 5854−5862. (39) Toschev, S.; Milchev, A.; Stoyanov, S. J. Cryst. Growth 1972, 13−14, 123−127. (40) Zumstein, R. C.; Rousseau, R. W. AIChE J. 1987, 33, 1921− 1925. (41) Zumstein, R. C.; Rousseau, R. W. AIChE J. 1987, 33, 121−129. (42) Ochsenbein, D. R.; Schorsch, S.; Salvatori, F.; Vetter, T.; Morari, M.; Mazzotti, M. Chem. Eng. Sci. 2015, 133, 30−43. (43) Mealey, D.; Zeglinski, J.; Khamar, D.; Rasmuson, Å. C. Faraday Discuss. 2015, 179, 309−328. (44) Snyder, D. L.; Miller, M. I. Random Point Processes in Time and Space; Springer Texts in Electrical Engineering; Springer-Verlag: New York, 1991. (45) Gibbs, J. W.; Bumstead, H. A.; Van Name, R. G.; Longley, W. R. The Collected Works of J. Willard Gibbs; Longmans, Green and Co., 1928. (46) Becker, R.; Döring, W. Ann. Phys. 1935, 416, 719−752. 6711

DOI: 10.1021/acs.cgd.7b01313 Cryst. Growth Des. 2017, 17, 6703−6711